Lucy Hicks

March 07, 2025

2180

Added to Email Alert

A new artificial intelligence (AI) tool can simultaneously assess blood samples for infections, autoimmune diseases, and vaccine responses, according to a new study.

The algorithm analyzed genes encoding immune cell receptors to identify patients with COVID-19, HIV, type 1 diabetes, lupus, those who recently received a flu vaccination, and healthy controls.

These receptor sequences are “unique sources of potential diagnostic information,” said Scott Boyd, MD, PhD, a professor of pathology at Stanford University, Stanford, California, and one of the senior authors of the paper. “They haven’t really been used to diagnose anything apart from cancers that are derived from B cells and T cells at present.”

While the study, published in Science on February 21, is a proof of concept, the AI tool could eventually be used to distinguish between diseases with overlapping symptoms, added study co-author Maxim Zaslavsky, PhD, a computer science postdoc at Stanford University.

Decoding the Language of the Immune System

These B-cell and T-cell receptors serve as a record of the body’s immune activity. If researchers could develop a method that effectively reads these responses, “we, in principle, would have an amazing diagnostic” that could identify multiple health conditions in just one test, said computational immunologist Ramy Arnaout, MD, DPhil, an associate professor of pathology at Harvard Medical School and associate director of the Clinical Microbiology Laboratories at Beth Israel Deaconess Medical Center in Boston. He was not involved with the work.

However, individual immune responses to the same pathogen can vary widely, Arnaout explained. While they may have similarities, it’s rare for two people with the same condition to have the same immune cell receptors.

“Traditional [sequencing] approaches sometimes struggle to group together the immune receptors that may look slightly different but actually recognize the same target,” Zaslavsky said. Large language models (LLMs) — the same underlying technology behind ChatGPT — “excel at this kind of task,” he continued.

“The idea here is that everybody’s repertoire of [B-cell and T-cell receptors] is like dialects of the same language. You say things slightly differently, and that’s reflected by these differences in sequence,” Arnaout said. By looking at enough immune sequences from different people with the same condition, LLMs can “figure out what the language is so that they get the dialect.”

Combining B-Cell and T-Cell Data

In the study, Boyd, Zaslavsky, and colleagues combined multiple AI models to analyze 23.5 million T-cell receptor sequences and 16.2 million B-cell receptor sequences in blood samples from 593 individuals. Across all individuals, 63 had COVID-19, 95 had HIV, 86 had lupus, 92 had type 1 diabetes, 37 recently received the flu vaccine, and 220 were healthy controls.

The algorithm — named Machine Learning for Immunological Diagnosis — most accurately screened samples that had both B-cell and T-cell receptor data. An analysis of the 542 samples that had this paired data achieved an area under the receiver operating characteristic curve (AUROC) of 0.986 and 85.3% accuracy.

Using B-cell receptors alone to screen patients, the AI tool achieved an AUROC of 0.959 and 74.0% accuracy. Analyzing T cells alone, the AUROC was 0.952 and was 75.1% accurate.

Previous analyses have largely focused on either B cells or T cells, Zaslavsky said, “but we saw clear evidence that combining both gives you a fuller panoramic picture of the immune activity because these two arms of the immune system are working together.”

B-cell receptor sequences were more informative in screening for COVID-19, HIV, and flu vaccination, while lupus and type 1 diabetes had clearer T-cell receptor signatures.

Clinical Applications

While a tool like this is far away from clinical use, “it’s an important step forward,” Arnaout said. A multiplex diagnostic test — a test that can detect multiple biomarkers from a single sample — could help simplify and shorten differential diagnoses, Arnaout added. As sequencing costs continue to come down, they could also be cost-saving as they could reduce the need for multiple, expensive lab tests.

These tests also have broader applications beyond diagnostics. Similar to how genomic sequencing helped advance targeted treatments for certain cancers, Zaslavsky hopes immune cell receptor sequencing can do the same for other areas of medicine. By scanning the immune system’s underlying activity, an AI tool like this could potentially help identify the most effective treatments for autoimmune disease, he said, and avoid trial and error.

“This is all just a research project for now, but we hope to see if this [approach] could follow in the footsteps of the way sequencing advances have transformed cancer [treatment].”

The study was funded by a variety of grants from various institutes of the National Institutes of Health and other institutional fellowships or awards. Boyd and Zaslavsky and five other authors are co-inventors of a patent related to this work. Boyd has consulted for Regeneron, Sanofi, Novartis, Genentech, Visterra, and Janssen Pharmaceuticals, and owns stock in AbCellera Biologics. Some other authors had a variety of financial relationships with pharmaceutical and biotechnology companies. Arnaout had no relevant disclosures.

2